Message/Author |
|
Dm posted on Monday, September 16, 2013 - 7:06 am
|
|
|
Hi, I have complex survey data with weights and stratification. I have a dichotomous binary dependent variable and missing on several IVs. I know I can use a probit regression but it is possible to run a logistic regression and use Montecarlo integration to avoid Listwise deletion. my input file is below: VARIABLE: NAMES ARE ID weight1 weight2 str SUBPOP anx12 sexf age educ Inc_LM SWD Nevermar Chronic; USEVARIABLES ARE anx12 sexf age educ Inc_LM SWD Nevermar Chronic; CATEGORICAL ARE anx12; MISSING ARE ALL (-99, -999); WEIGHT = WEIGHT2; STRATIFICATION IS STR; IDVARIABLE=ID; SUBPOPULATION = SUBPOP EQ 1; ANALYSIS: Type = Complex; Estimator = ; ALGORITHM = INTEGRATION; integration = montecarlo; MODEL: anx12 ON sexf age educ Inc_LM SWD Nevermar chronic; [sexf age educ Inc_LM SWD Nevermar chronic]; output: TECH1 TECH4 ; standardized; sampstat; CINTERVAL; |
|
|
That should work. I would use INTEGRATION = MONTECARLO (5000); |
|
|
I generated factor scores using approach in Example 11.7, using NDATASETS=50. I also saved plausible values along with mean, median, etc. in separate file (Thanks for pointing me to that example, Linda...). I then analyzed latent structural model based on those factor scores, using Example 13.13 method. (All variables are continuous.). There was a considerable difference between ML and MLR estimators in terms of test of fit, RMSEA, CFI/TLI, and SRMR: ML Estimator: Chi2=14.28, 23 DF; RMSEA=0.000; CFI=1.000; TLI=1.134 MLR Estimator (Means): Chi2=453.13, 23DF; RMSEA=0.237; CFI=0.822; TLI=0.659; SRMR=0.072. All I did was change ESTIMATOR=MLR to ESTIMATOR=ML; everything else is comparable, though only MLR provides me with means & SD for fit statistic estimates. Which results are more appropriate? Is there a different approach that I should use? Thanks in advance... (BTW: I'm using version 7.11) |
|
|
Please send the two outputs and your license number to support@statmodel.com. |
|
|
Done - thanks... Also (is probably a naïve question), is there a way to incorporate the Bayes-generated plausible factor score values and distribution information (as per the Example 11.7 SAVEDATA: file=ex11.7plaus.dat ...) into an analysis testing relationships among those latent variables using a Bayes estimator? Or is my use of the imputation approach (as described in previous post) essentially the same thing, so I have basically already done that? Thanks! Michael |
|
|
You can't use the Bayes estimator with TYPE=IMPUTATION. |
|
|
Thanks. So, to be sure I understand - the factor scores generated in the imputed data sets in the original measurement model, using Bayes estimator, reflect the variability in the values, so that the analysis using the imputation data sets incorporates / reflects the error variance for those latent variable values estimated in the original measurement model. Correct? Sorry about being a bit slow on the uptake... Michael |
|
|
Yes, that is correct. That's the advantage of Bayes plausible values. |
|
db40 posted on Sunday, October 06, 2013 - 9:12 am
|
|
|
Professor Muthen, regarding the post at the top of this page posted by DM. my question is: Is monte carlo integration an appropriate technique for use with with missing on categorical covariates? |
|
|
Monte Carlo integration is a special type of algorithm that is required when there is missing data on a mediator. It is not related to missing data estimation. |
|
Back to top |